Merge pull request #1204 from heshpdx:master

PiperOrigin-RevId: 697922880
This commit is contained in:
Copybara-Service 2024-11-19 01:42:21 -08:00
commit 2b6efcbdcc
3 changed files with 10 additions and 5 deletions

View File

@ -118,6 +118,8 @@ static size_t FN(FindBlocks)(const DataType* data, const size_t length,
size_t insert_cost_ix = symbol * num_histograms;
double min_cost = 1e99;
double block_switch_cost = block_switch_bitcost;
static const size_t prologue_length = 2000;
static const double multiplier = 0.07 / 2000;
size_t k;
for (k = 0; k < num_histograms; ++k) {
/* We are coding the symbol with entropy code k. */
@ -128,8 +130,8 @@ static size_t FN(FindBlocks)(const DataType* data, const size_t length,
}
}
/* More blocks for the beginning. */
if (byte_ix < 2000) {
block_switch_cost *= 0.77 + 0.07 * (double)byte_ix / 2000;
if (byte_ix < prologue_length) {
block_switch_cost *= 0.77 + multiplier * (double)byte_ix;
}
for (k = 0; k < num_histograms; ++k) {
cost[k] -= min_cost;

View File

@ -437,9 +437,10 @@ static BROTLI_BOOL ShouldCompress(
if ((double)num_literals > 0.99 * (double)bytes) {
uint32_t literal_histo[256] = { 0 };
static const uint32_t kSampleRate = 13;
static const double kInvSampleRate = 1.0 / 13.0;
static const double kMinEntropy = 7.92;
const double bit_cost_threshold =
(double)bytes * kMinEntropy / kSampleRate;
(double)bytes * kMinEntropy * kInvSampleRate;
size_t t = (bytes + kSampleRate - 1) / kSampleRate;
uint32_t pos = (uint32_t)last_flush_pos;
size_t i;

View File

@ -106,6 +106,8 @@ static void EstimateBitCostsForLiteralsUTF8(size_t pos, size_t len, size_t mask,
size_t utf8_pos = UTF8Position(last_c, c, max_utf8);
size_t masked_pos = (pos + i) & mask;
size_t histo = histogram[256 * utf8_pos + data[masked_pos]];
static const size_t prologue_length = 2000;
static const double multiplier = 0.35 / 2000;
double lit_cost;
if (histo == 0) {
histo = 1;
@ -120,8 +122,8 @@ static void EstimateBitCostsForLiteralsUTF8(size_t pos, size_t len, size_t mask,
Perhaps because the entropy source is changing its properties
rapidly in the beginning of the file, perhaps because the beginning
of the data is a statistical "anomaly". */
if (i < 2000) {
lit_cost += 0.7 - ((double)(2000 - i) / 2000.0 * 0.35);
if (i < prologue_length) {
lit_cost += 0.35 + multiplier * (double)i;
}
cost[i] = (float)lit_cost;
}